Embeddings: the foundation of our on-device assistant
Embeddings are the foundation of LLMs, RAG, and our on-device AI assistant at CES
- A great way to understand them is through an interactive visualization.
To prepare data for the assistant, we converted a user-manual PDF into text chunks, created structured question-answer pairs, and transformed those text questions into vectors using an embedding model.
This is a very similar format to the vector databases used for Retrieval Augmented Generation (RAG). Our vectors have over 300 dimensions, reduced to just 3 for the visualization. These vectors represent meaning, which enables semantic search of voice queries, so you don't need to phrase commands perfectly. It is literally looking for the closest match in semantic space.